Is AMD killing the nascent ARM server market?

By: Linus Torvalds (torvalds.delete@this.linux-foundation.org), June 8, 2018 9:56 am
Room: Moderated Discussions
RichardC (tich.delete@this.pobox.com) on June 8, 2018 7:49 am wrote:
>
> They also care a great deal about physical density (cpu throughput per rack) and
> power efficiency (cpu throughput per watt). If ARM provides a significantly better
> solution on those dimensions for some parts of their workload, it can be successful.

This whole constant fixation with power efficiency of just the inner core needs to stop.

In a data center, DRAM power use tends to be on the same size as CPU power use. It obviously matters hugely what the load is and what kind of server you're talking of (some servers tend to be much more memory-oriented, others care way more about pure CPU power).

But people really seem to forget that the CPU is just one component, and very seldom even the dominant one. All the "ARM is so power efficient" people continually seem to ignore it, whether it's in laptops (screen tends to dominate) or when they are in servers (memoryis a big deal, even if not necessarily dominant, but there's also networking and storage).

So it's actually fundamentally hard to provide a "significantly better solution" with instruction set changes. Because you're only attacking a fraction of the power use.

Here's an old paper

Data center energy consumption modeling

and I don't guarantee how accurate it is today, but see Fig 7: it estimates that 33% of the server power use was due to CPU's for google in 2007.

NOTE! That's all of the CPU - including things like caches etc that have nothing to do with architecture. And it's still just a third of the server power use. The actual instruction set effect is much less than that.

And then if you look at the big picture, you'll see that that "server power use" in turn is actually only a third of the data center power use, with lighting, cooling, power conversion etc being the dominant costs.

And yes, some of cooling obviously comes from the CPU too (but not all, don't make that mistake either), so this isn't all entirely independent, but Christ people! You need to stop thinking of "instruction set power use". It's such a small part that it's almost not going to be noticeable.

If you make your CPU core three times more power efficient, what does that do to the power efficiency of your caches? Of your memory? Of your networking? Of your lighting and cooling and power delivery? Not a whole lot. It might nudge some of the others, but it's not going to be the revolutionary change that some people seem to think it is.

And that's when it's some completely radical "three times more power efficient" thing. When you actually get to high-performance ARM, that's not what people are talking about any more. They are talking about "tens of percent" kind of advantages.

So give it up already. The power efficiency thing was always a bedtime story for retarded squirrels.

If ARM wants to take over the world, they will do it with volume and low price and the flexibility of having lots of manufacturers to fill all the niches (although that last one right now is probably a negative, because in the server space it means "fragmentation and lack of good platforms").

Power efficiency is a second-order effect at best.

Linus
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
Is AMD killing the nascent ARM server market?Doug S2018/06/08 01:41 AM
  Is AMD killing the nascent ARM server market?dmcq2018/06/08 02:22 AM
    Is AMD killing the nascent ARM server market?Brendan2018/06/08 02:41 AM
      Is AMD killing the nascent ARM server market?dmcq2018/06/08 03:01 AM
        Is AMD killing the nascent ARM server market?dmcq2018/06/08 03:09 AM
          Is AMD killing the nascent ARM server market?Jukka Larja2018/06/09 05:25 AM
    Is AMD killing the nascent ARM server market?Aaron Spink2018/06/08 03:33 AM
      Is AMD killing the nascent ARM server market?dmcq2018/06/08 04:09 AM
        Is AMD killing the nascent ARM server market?anon.12018/06/08 05:59 PM
      Is AMD killing the nascent ARM server market?RichardC2018/06/08 07:49 AM
        Is AMD killing the nascent ARM server market?Michael S2018/06/08 08:24 AM
        Is AMD killing the nascent ARM server market?Linus Torvalds2018/06/08 09:56 AM
          Is AMD killing the nascent ARM server market?Yuhong Bao2018/06/08 11:59 AM
          Is AMD killing the nascent ARM server market?RichardC2018/06/08 12:40 PM
          Is AMD killing the nascent ARM server market?wumpus2018/06/08 08:31 PM
        Is AMD killing the nascent ARM server market?Aaron Spink2018/06/08 05:57 PM
          Is AMD killing the nascent ARM server market?dmcq2018/06/09 06:02 AM
    Intel 28c@5GHz """HEDT"""Jeff S.2018/06/08 08:31 AM
      Intel 28c@5GHz """HEDT"""Eric Bron2018/06/08 09:37 AM
        Intel 28c@5GHz """HEDT"""Eric Bron2018/06/08 09:43 AM
        Intel 28c@5GHz """HEDT"""Jeff S.2018/06/08 10:57 AM
          Intel 28c@5GHz """HEDT"""Eric Bron2018/06/08 03:16 PM
            Intel 28c@5GHz """HEDT"""Aaron Spink2018/06/08 06:19 PM
              Intel 28c@5GHz """HEDT"""Eric Bron2018/06/08 11:34 PM
        Intel 28c@5GHz """HEDT"""Doug S2018/06/08 03:46 PM
          Intel 28c@5GHz """HEDT"""Eric Bron2018/06/08 03:52 PM
          Intel 28c@5GHz """HEDT"""juanrga2018/06/08 05:51 PM
          Intel 28c@5GHz """HEDT"""Aaron Spink2018/06/08 06:23 PM
            Intel 28c@5GHz """HEDT"""Jeff S.2018/06/08 09:33 PM
              Intel 28c@5GHz """HEDT"""Aaron Spink2018/06/09 01:42 AM
                Intel 28c@5GHz """HEDT"""Jeff S.2018/06/09 11:36 AM
                  Intel 28c@5GHz """HEDT"""pgerassi2018/06/09 04:11 PM
            Intel 28c@5GHz """HEDT"""hobold2018/06/09 12:24 AM
              Intel 28c@5GHz """HEDT"""wumpus2018/06/09 07:08 PM
      Intel 28c@5GHz """HEDT"""juanrga2018/06/08 10:48 AM
      Intel 28c@5GHz """HEDT"""Aaron Spink2018/06/08 06:04 PM
      Intel 28c@5GHz """HEDT"""Groo2018/06/08 06:20 PM
        Intel 28c@5GHz """HEDT"""Jeff S.2018/06/08 10:17 PM
          Intel 28c@5GHz """HEDT"""hobold2018/06/09 12:38 AM
            Intel 28c@5GHz """HEDT"""juanrga2018/06/09 11:00 AM
              Intel 28c@5GHz """HEDT"""hobold2018/06/09 01:11 PM
                Intel 28c@5GHz """HEDT"""juanrga2018/06/10 03:36 AM
                  TDP != power consumptionhobold2018/06/10 04:46 AM
                    TDP != power consumption (we know that, thanks)juanrga2018/06/10 06:02 AM
                      TDP != power consumption (we know that, thanks)hobold2018/06/10 07:11 AM
                Intel 28c@5GHz """HEDT"""Bigos2018/06/10 03:41 AM
          Intel 28c@5GHz """HEDT"""juanrga2018/06/09 10:48 AM
            Intel 28c@5GHz """HEDT"""Jeff S.2018/06/09 11:52 AM
  Zen power/performanceMichael S2018/06/08 04:57 AM
    Zen power/performanceDoug S2018/06/08 09:31 AM
      Zen power/performanceMichael S2018/06/08 09:35 AM
  Is AMD killing the nascent ARM server market?juanrga2018/06/08 10:12 AM
    rainbow unicorn pegasus samplehobold2018/06/08 01:45 PM
      rainbow unicorn pegasus samplejokerman2018/06/08 02:54 PM
        rainbow unicorn pegasus samplehobold2018/06/08 03:55 PM
      rainbow unicorn pegasus samplejuanrga2018/06/08 06:06 PM
        rainbow unicorn pegasus sampleJeff S.2018/06/08 10:02 PM
          rainbow unicorn pegasus samplejuanrga2018/06/09 11:04 AM
        rainbow unicorn pegasus samplehobold2018/06/09 12:11 AM
          rainbow unicorn pegasus samplejuanrga2018/06/09 11:21 AM
            rainbow unicorn pegasus samplehobold2018/06/09 01:38 PM
      rainbow unicorn pegasus sampleAaron Spink2018/06/08 06:28 PM
        rainbow unicorn pegasus samplehobold2018/06/08 11:32 PM
          rainbow unicorn pegasus sampleAdrian2018/06/09 12:09 AM
            rainbow unicorn pegasus samplehobold2018/06/09 12:47 AM
            rainbow unicorn pegasus samplejuanrga2018/06/09 11:30 AM
              rainbow unicorn pegasus sampleDoug S2018/06/09 11:53 AM
                rainbow unicorn pegasus samplewumpus2018/06/09 06:44 PM
              rainbow unicorn pegasus sampleAdrian2018/06/09 12:59 PM
                rainbow unicorn pegasus samplejuanrga2018/06/10 04:06 AM
      Core performance isn't independent; it's highly correlatedCarl L2018/06/08 07:52 PM
        rainbow unicorn pegasus processwumpus2018/06/08 08:54 PM
        Core performance isn't independent; it's highly correlatedhobold2018/06/08 11:57 PM
          Core performance isn't independent; it's highly correlatedAaron Spink2018/06/09 01:54 AM
            Core performance isn't independent; it's highly correlatedhobold2018/06/09 03:41 AM
              Core performance isn't independent; it's highly correlatedDoug S2018/06/09 11:55 AM
                Core performance isn't independent; it's highly correlatedhobold2018/06/09 01:50 PM
            don't hold your breath for a 5GHz productwumpus2018/06/09 06:56 PM
              don't hold your breath for a 5GHz productEric Bron2018/06/10 01:02 AM
                don't hold your breath for a 5GHz producthobold2018/06/10 01:34 AM
                  don't hold your breath for a 5GHz productjuanrga2018/06/10 03:24 AM
                    don't hold your breath for a 5GHz producthobold2018/06/10 04:12 AM
                don't hold your breath for a 5GHz productjuanrga2018/06/10 03:15 AM
              don't hold your breath for a 5GHz productjuanrga2018/06/10 03:14 AM
                don't hold your breath for a 5GHz productMichael_S2018/06/10 05:33 AM
                  don't hold your breath for a 5GHz productMichael_S2018/06/10 05:34 AM
                    don't hold your breath for a 5GHz productbakaneko2018/06/10 05:49 AM
                  don't hold your breath for a 5GHz productjuanrga2018/06/10 06:31 AM
              don't hold your breath for a 5GHz productMichael_S2018/06/10 03:26 AM
          Core performance isn't independent; it's highly correlatedCarl L2018/06/10 12:49 AM
            Core performance isn't independent; it's highly correlatedhobold2018/06/10 01:24 AM
              wow you are so smartjokerman2018/06/10 01:28 AM
                you can be smart, too!hobold2018/06/10 04:02 AM
              Core performance isn't independent; it's highly correlatedbakaneko2018/06/10 03:49 AM
                Core performance isn't independent; it's highly correlatedhobold2018/06/10 04:09 AM
                  Core performance isn't independent; it's highly correlatedbakaneko2018/06/10 05:39 AM
                    Stupidity Limit Theoremhobold2018/06/10 07:21 AM
                      Stupidity Limit Theorembakaneko2018/06/10 10:28 AM
                        Modelling Limitshobold2018/06/10 12:08 PM
                          Modelling Limitsbakaneko2018/06/11 02:40 AM
                  Core performance isn't independent; it's highly correlatedMaynard Handley2018/06/10 01:17 PM
              Process variationDavid Kanter2018/06/10 08:31 AM
                Process variationhobold2018/06/10 11:55 AM
                  Process variationMaynard Handley2018/06/10 01:33 PM
                  Process variationanon2018/06/10 02:57 PM
                    Process variationhobold2018/06/11 11:42 AM
                      Process variationanon2018/06/11 12:11 PM
                    Well said in a thread of mostly nonsense (NT)someone2018/06/13 08:44 AM
              Core performance isn't independent; it's highly correlatedCarl L2018/06/10 11:31 AM
                Core performance isn't independent; it's highly correlatedhobold2018/06/10 12:38 PM
                  Core performance isn't independent; it's highly correlatedDoug S2018/06/10 12:46 PM
                  Core performance isn't independent; it's highly correlatedAdrian2018/06/10 02:08 PM
                    Core performance isn't independent; it's highly correlatedhobold2018/06/11 12:44 PM
                      Core performance isn't independent; it's highly correlatedDoug S2018/06/11 03:44 PM
                        Core performance isn't independent; it's highly correlatedhobold2018/06/12 06:15 AM
                          Core performance isn't independent; it's highly correlatedanon2018/06/12 05:46 PM
                            Core performance isn't independent; it's highly correlatedhobold2018/06/14 06:26 AM
                              Core performance isn't independent; it's highly correlatedsomeone2018/06/14 06:58 AM
                              Core performance isn't independent; it's highly correlatedanon2018/06/14 04:45 PM
                      Core performance isn't independent; it's highly correlatedRicardo B2018/06/12 06:51 AM
        Core performance isn't independent; it's highly correlatedanon2018/06/09 09:04 PM
          Core performance isn't independent; it's highly correlatedJukka Larja2018/06/09 11:56 PM
      rainbow unicorn pegasus samplebakaneko2018/06/09 05:35 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell green?